Meet the team that turns your ideas into work.

Discover EightBots
EIGHTBOTS
Contact salesSign up
AI provider

Kyutai

Kyutai is an independent nonprofit AI research lab in Paris. It develops language, speech and visual models and publishes research code and model releases with terms that vary by project.

Explore models

Models & tools

  • A text encoder with fourfold pooling trained for use with a Llama decoder.

    Launch date not confirmedResearch releaseDownloadable

  • A text encoder with eightfold pooling trained for the Llama 3.1 8B decoder.

    Launch date not confirmedResearch releaseDownloadable

  • A Llama-derived text encoder with eightfold pooling trained for the Mistral 7B decoder.

    Launch date not confirmedResearch releaseDownloadable

  • A text encoder with eightfold pooling trained with both Llama 3.1 8B and Mistral 7B decoders.

    Launch date not confirmedResearch releaseDownloadable

  • A causal video autoencoder with one-frame temporal compression and a 16-by-16-pixel spatial grid.

    Launch date not confirmedResearch releaseDownloadable

  • A causal video autoencoder with one-frame temporal compression and a 48-by-48-pixel spatial grid.

    Launch date not confirmedResearch releaseDownloadable

  • A causal video autoencoder with two-frame temporal compression and a 16-by-16-pixel spatial grid.

    Launch date not confirmedResearch releaseDownloadable

  • Four separately trained world models that predict video from context frames and actions in distinct environments.

    Launch date not confirmedResearch releaseDownloadable

  • A multi-resolution CADRE autoencoder with two-frame and 16-by-16-pixel compression.

    Launch date not confirmedResearch releaseDownloadable

  • A causal video autoencoder with two-frame temporal compression and a 32-by-32-pixel spatial grid.

    Launch date not confirmedResearch releaseDownloadable

  • A multi-resolution CADRE autoencoder with two-frame and 32-by-32-pixel compression.

    Launch date not confirmedResearch releaseDownloadable

  • A causal video autoencoder with four-frame temporal compression and a 16-by-16-pixel spatial grid.

    Launch date not confirmedResearch releaseDownloadable

  • A multi-resolution CADRE autoencoder with four-frame and 16-by-16-pixel compression.

    Launch date not confirmedResearch releaseDownloadable

  • A causal video autoencoder with four-frame temporal compression and a 24-by-24-pixel spatial grid.

    Launch date not confirmedResearch releaseDownloadable

  • A causal video autoencoder with eight-frame temporal compression and a 16-by-16-pixel spatial grid.

    Launch date not confirmedResearch releaseDownloadable

  • A multi-resolution CADRE autoencoder with eight-frame and 16-by-16-pixel compression.

    Launch date not confirmedResearch releaseDownloadable

  • Four inverse-dynamics models that estimate actions from generated frames to evaluate CADRE world models.

    Launch date not confirmedResearch releaseDownloadable

  • A vision-language model that links Helium 1 with a Qwen 2.5 VL encoder through cross-attention.

    Launch date not confirmedResearch releaseDownloadable

  • A Helium vision-language model that shares weights between self-attention and cross-attention.

    Launch date not confirmedResearch releaseDownloadable

  • A Qwen 2.5 VL 3B Instruct adaptation that replaces image-token insertion with cross-attention.

    Launch date not confirmedResearch releaseDownloadable

  • A CASA Qwen checkpoint fine-tuned on LiveWhisperX-526K for live video commentary.

    Launch date not confirmedResearch releaseDownloadable

  • A Qwen-based vision-language adaptation with shared self-attention and cross-attention weights.

    Launch date not confirmedResearch releaseDownloadable

  • A set of fastText classifiers that filters Common Crawl text in the 24 official EU languages.

    Launch date not confirmedResearch releaseDownloadable

  • A GLM 4 Voice 9B adaptation trained with reinforcement learning for spoken mathematical reasoning.

    Launch date not confirmedResearch releaseDownloadable

  • A GLM 4 Voice 9B adaptation that combines silent text reasoning with spoken output and reinforcement learning.

    Launch date not confirmedResearch releaseDownloadable

  • A 2B base language model for the 24 official EU languages, built by combining selected trained variants.

    Launch date not confirmedResearch releaseDownloadable

  • A 2B Helium base checkpoint trained on the textbook-oriented part of its filtered data.

    Launch date not confirmedResearch releaseDownloadable

  • A 2B Helium base checkpoint trained on the humanities-oriented part of its filtered data.

    Launch date not confirmedResearch releaseDownloadable

  • A 2B Helium base checkpoint from the life-oriented part of its modular training study.

    Launch date not confirmedResearch releaseDownloadable

  • The main 2B Helium training checkpoint, separate from the final model that combines trained variants.

    Launch date not confirmedResearch releaseDownloadable

  • A 2B Helium base checkpoint trained on pop-culture-oriented data.

    Launch date not confirmedResearch releaseDownloadable

  • A 2B Helium base checkpoint trained on scientific text selected by its data filter.

    Launch date not confirmedResearch releaseDownloadable

  • A 2B Helium base checkpoint for science, technology, engineering and mathematics training data.

    Launch date not confirmedResearch releaseDownloadable

  • A 2B Helium base checkpoint trained on Wikipedia-oriented content.

    Launch date not confirmedResearch releaseDownloadable

  • A 2B base language model for English, French, German, Italian, Portuguese and Spanish.

    Launch date not confirmedResearch releaseDownloadable

  • An instruction-tuned vision-language reference model that inserts image tokens into the text sequence.

    Launch date not confirmedResearch releaseDownloadable
  • Hibiki 1B

    Kyutai · Hibiki

    A 1.7B simultaneous French-to-English speech translation model that aims to preserve the speaker's voice.

    Launch date not confirmedResearch releaseDownloadable
  • Hibiki 2B

    Kyutai · Hibiki

    A 2.7B simultaneous French-to-English speech translation model that aims to preserve the speaker's voice.

    Launch date not confirmedResearch releaseDownloadable

  • A speech model that translates French, Spanish, Portuguese and German into English using sentence-aligned training.

    Launch date not confirmedResearch releaseDownloadable

  • A streaming English and French transcription model with a configured output delay of 0.5 seconds.

    Launch date not confirmedResearch releaseDownloadable

  • A streaming English transcription model with a configured output delay of 2.5 seconds.

    Launch date not confirmedResearch releaseDownloadable

  • An English streaming speech model trained on public data, using 16 audio tokens per frame.

    Launch date not confirmedResearch releaseDownloadable

  • An English and French streaming speech model with voice conditioning and adjustable audio token count.

    Launch date not confirmedResearch releaseDownloadable
  • Mimi

    Kyutai · Mimi

    A streaming neural audio codec that compresses speech into tokens and reconstructs sound from them.

    Launch date not confirmedResearch releaseDownloadable
  • MIRA

    Kyutai · MIRA

    A joint General Intuition and Kyutai world model that generates four synchronized Rocket League views from player actions.

    Launch date not confirmedResearch releaseProvider product

  • A Moshi speech-dialogue model that listens while speaking, using Kyutai's Moshika voice.

    Launch date not confirmedResearch releaseDownloadable

  • A Moshi speech-dialogue model that listens while speaking, using Kyutai's Moshiko voice.

    Launch date not confirmedResearch releaseDownloadable
  • Moshika RL Seamless

    Kyutai · Interactivity alignment

    A Moshika speech-dialogue adaptation trained with reinforcement learning for pauses, turn-taking and interruptions.

    Launch date not confirmedRestricted accessDownloadable

  • A Moshika speech front end that receives retrieved text while an audio conversation continues.

    Launch date not confirmedResearch releaseDownloadable

  • A Moshika adaptation that connects spoken dialogue to image features through trained cross-attention.

    Launch date not confirmedResearch releaseDownloadable

  • A large Kyutai and Mirelo model that transcribes recorded music into multi-instrument MIDI.

    Launch date not confirmedRestricted accessDownloadable

  • A medium-sized Kyutai and Mirelo model that transcribes recorded music into multi-instrument MIDI.

    Launch date not confirmedRestricted accessDownloadable

  • A small Kyutai and Mirelo model that transcribes recorded music into multi-instrument MIDI.

    Launch date not confirmedRestricted accessDownloadable
  • OVIE

    Kyutai · OVIE

    A single-image model that generates new scene views from a target camera pose at 256-by-256 resolution.

    Launch date not confirmedResearch releaseDownloadable

  • An OVIE model retrained at 512-by-512 resolution for single-image novel-view generation.

    Launch date not confirmedResearch releaseDownloadable

  • An OVIE novel-view model fine-tuned for 50,000 steps on paired DL3DV multi-view data.

    Launch date not confirmedResearch releaseDownloadable

  • An OVIE novel-view model fine-tuned for 50,000 steps on paired RealEstate10K indoor scenes.

    Launch date not confirmedResearch releaseDownloadable
  • PersonaPlex RL Seamless

    Kyutai · Interactivity alignment

    A NVIDIA PersonaPlex 7B v1 adaptation trained with reinforcement learning for conversational timing.

    Launch date not confirmedRestricted accessDownloadable

  • A Dutch text-to-speech model with a six-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A Dutch speech-generation checkpoint with a 24-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A distilled English speech model with 1.9M parameters including its decoder.

    Launch date not confirmedResearch releaseDownloadable

  • A 1.9M-parameter English speech model with reward fine-tuning after distillation.

    Launch date not confirmedResearch releaseDownloadable

  • An 18.2M-parameter distilled English speech model with an eight-layer backbone and a reduced decoder.

    Launch date not confirmedDownloadable

  • An 18.2M-parameter English speech model with reward fine-tuning added after distillation.

    Launch date not confirmedDownloadable

  • A January 2026 English speech checkpoint with a six-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • An April 2026 English speech checkpoint with a six-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A September 2026 English speech checkpoint with a six-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A September 2026 English speech revision distilled on re-aligned Dora data with moving-average weights.

    Launch date not confirmedDownloadable

  • An October 2026 English speech checkpoint that retrains September revision 2 with a corrected training loader.

    Launch date not confirmedDownloadable

  • An April 2026 English speech checkpoint with a 24-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A September 2026 English speech checkpoint with a 24-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • An experimental English speech model that trains its sampler head with a drifting objective.

    Launch date not confirmedResearch releaseDownloadable

  • A distilled English speech model with 6.9M parameters including its decoder.

    Launch date not confirmedDownloadable

  • A 6.9M-parameter English speech model with reward fine-tuning after distillation.

    Launch date not confirmedDownloadable

  • A French text-to-speech model with a six-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A French text-to-speech model with a 24-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A German text-to-speech model with a six-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A German text-to-speech model with a 24-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • An Italian text-to-speech model with a six-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • An Italian text-to-speech model with a 24-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A Portuguese text-to-speech model with a six-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A Portuguese text-to-speech model with a 24-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A Spanish text-to-speech model with a six-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A Spanish text-to-speech model with a 24-layer Pocket TTS backbone.

    Launch date not confirmedDownloadable

  • A Qwen-based vision-language-action model pretrained on robot, simulation, human-hand and vision-language data.

    Launch date not confirmedResearch releaseDownloadable

  • A REKI action policy for the simulated Franka Panda robot in the four LIBERO evaluation suites.

    Launch date not confirmedResearch releaseDownloadable

  • A REKI kitchen-task policy for a simulated Franka Panda arm on an Omron mobile base.

    Launch date not confirmedResearch releaseDownloadable

  • A dual-arm AgileX ALOHA policy trained on the clean scenes of 50 RoboTwin tasks.

    Launch date not confirmedResearch releaseDownloadable

  • A dual-arm REKI policy trained on both clean and randomized RoboTwin scenes.

    Launch date not confirmedResearch releaseDownloadable

  • A 6B research model family comparing chronological and shuffled text training for changing facts.

    Launch date not confirmedResearch releaseDownloadable
  • Surflo v0

    Kyutai · Surflo

    A multi-view reconstruction model that generates oriented surface points from RGB images without supplied camera poses.

    Launch date not confirmedResearch releaseDownloadable